Machine learning in Madrid (zoom)
Lunes, 14 de febrero de 2022, 11:30-12:30h (horario diferente al habitual!!!!)
Ponente: Vivak Patel (UW-Madison)
Título: Global Convergence and Stability of Stochastic Gradient Descent
Abstract: Stochastic gradient descent (SGD) is widely deployed in a number of different disciplines, often on non-convex problems with complicated noise models. However, SGD's existing convergence theory often does not apply. In this talk, we will establish the need for a convergence theory under broader assumptions with some simple examples. We will then state a global convergence result for SGD under these broad assumptions. Then, we will discuss the issue of stability, which addresses what happens when SGD's iterates diverge.